Download The Game World is Flat: The Complexity of Nash Equilibria in

The Game World is Flat:
The Complexity of Nash Equilibria in Succinct Games
Constantinos Daskalakis† , Alex Fabrikant‡, and Christos H. Papadimitriou§
UC Berkeley, Computer Science Division
Abstract. A recent sequence of results established that computing Nash equilibria in normal form games is a PPAD-complete problem even in the case of two
players [11, 6, 4]. By extending these techniques we prove a general theorem,
showing that, for a far more general class of families of succinctly representable
multiplayer games, the Nash equilibrium problem can also be reduced to the twoplayer case. In view of empirically successful algorithms available for this problem, this is in essence a positive result — even though, due to the complexity of
the reductions, it is of no immediate practical significance. We further extend this
conclusion to extensive form games and network congestion games, two classes
which do not fall into the same succinct representation framework, and for which
no positive algorithmic result had been known.
1 Introduction
Nash proved in 1951 that every game has a mixed Nash equilibrium [15]. However, the
complexity of the computational problem of finding such an equilibrium had remained
open for more than half century, attacked with increased intensity over the past decades.
This question was resolved recently, when it was established that the problem is PPADcomplete [6] (the appropriate complexity level, defined in [18]) and thus presumably
intractable, for the case of 4 players; this was subsequently improved to three players
[5, 3] and, most remarkably, two players [4].
In particular, the combined results of [11, 6, 4] establish that the general Nash equilibrium problem for normal form games (the standard and most explicit representation)
and for graphical agames (an important succinct representation, see the next paragraph)
can all be reduced to 2-player games. 2-player games in turn can be solved by several
techniques such as the Lemke-Howson algorithm [14, 20], a simplex-like technique that
is known empirically to behave well even though exponential counterexamples do exist
[19]. In this paper we extend these results to essentially all known kinds of succinct
representations of games, as well as to more sophisticated concepts of equilibrium.
Besides this significant increase in our understanding of complexity issues, computational considerations also led to much interest in succinct representations of games.
Computer scientists became interested in games because they help model networks and
†
Supported by NSF ITR Grant CCR-0121555. [email protected]
Supported by the Fannie and John Hertz Foundation. [email protected]
§
Supported by NSF ITR CCR-0121555 grant and a Microsoft Research grant.
[email protected]
‡
auctions; thus we should mainly focus on games with many players. However, multiplayer games in normal form require in order to be described an amount of data that is
exponential in the number of players. When the number of players is large, the resulting
computational problems are hardly legitimate, and complexity issues are hopelessly distorted. This has led the community to consider broad classes of succinctly representable
games, some of which had been studied by traditional game theory for decades, while
others (like the graphical games [13]) were invented by computer scientists spurred by
the motivations outline above. (We formally define succinct games in the next section,
but also deal in this paper with two cases, network congestion games and extensive form
games, that do not fit within this definition.)
The first general positive algorithmic result for succinct games was obtained only
recently [17]: a polynomial-time algorithm for finding a correlated equilibrium (an important generalization of the Nash equilibrium due to Aumann [1]). The main result in
[17] states that a family of succinct games has a polynomial-time algorithm for correlated equilibria provided that there is a polynomial time oracle which, given a strategy
profile, computes the expected utility of each player.
In this paper, using completely different techniques inspired from [11], we show a
general result (Theorem 2) that is remarkably parallel to that of [17]: The Nash equilibrium problem of a family of succinct games can be reduced to the 2-player case provided
that a (slightly constrained) polynomial-length straight-line arithmetic program exists
which computes, again, the expected utility of a given strategy profile (notice the extra algebraic requirement here, necessitated by the algebraic nature of our techniques).
We proceed to point out that for all major known families of succinct games such a
straight-line program exists (Corollary 1).
We also extend these techniques to two other game classes, Network congestion
games [7] and extensive form games, which do not fit into our succinctness framework,
because the number of strategies is exponential in the input, and for which the result of
[17] does not apply, Theorems 3 and 4, respectively).
2 Definitions and Background
In a game in normal form we have r ≥ 2 players (and for each player p ≤ r a finite set
Sp of pure strategies. We denote the Cartesian product of the Sp ’s by S (the set of pure
strategy profiles) and the Cartesian product of the pure strategy sets of players other
than p by S−p . Finally, for each p ≤ r and s ∈ S we have a payoff ups .
A mixed strategy for player p is a distribution on Sp , that is, |Sp | nonnegative real
numbers adding to 1. Call a set of r mixed strategies xpj , p = 1, . . . , r, j ∈ Sp a Nash
Qr
P
equilibrium if, for each p, its expected payoff, s∈S ups q=1 xqsq is maximized over
all mixed strategies of p. That is, a Nash equilibrium is a set of mixed Q
strategies from
which no player has an incentive to deviate. For s ∈ S−p , let xs = q6=p xqsq . It is
well-known (see, e.g., [16]) that the following is an equivalent condition for a set of
mixed strategies to be a Nash equilibrium:
X p
X p
∀p, j
ujs xs >
uj 0 s xs =⇒ xpj0 = 0.
(1)
s∈S−p
s∈S−p
Also, a set of mixed strategies is an ε-Nash equilibrium for some ε > 0 if the following
holds:
X p
X p
ujs xs >
uj 0 s xs + ε =⇒ xpj0 = 0.
(2)
s∈S−p
s∈S−p
We next define the complexity class PPAD. An FNP search problem P is a set of
inputs IP ⊆ Σ ∗ such that for each x ∈ IP there is an associated set of solutions
k
k
Px ⊆ Σ |x| for some integer k, such that for each x ∈ IP and y ∈ Σ |x| whether
y ∈ Px is decidable in polynomial time (notice that this is precisely NP with an added
emphasis on finding a witness). For example, r-NASH is the search problem P in which
each x ∈ IP is an r-player game in normal form together with a binary integer A (the
1
-Nash equilibria of the game.
accuracy specification), and Px is the set of A
A search problem is total if Px 6= ∅ for all x ∈ IP . For example, Nash’s 1951
theorem [15] implies that r-NASH is total. The set of all total FNP search problems is
denoted TFNP. TFNP seems to have no generic complete problem, and so we study its
subclasses: PLS [12], PPP, PPA and PPAD [18]. In particular, PPAD is the class of all
total search problems reducible to the following:
END OF THE LINE : Given two circuits S and P with n input bits and n output bits,
such that P (0n ) = 0n 6= S(0n ), find an input x ∈ {0, 1}n such that P (S(x)) 6= x or
S(P (x)) 6= x 6= 0n .
Intuitively, END OF THE LINE creates a directed graph with vertex set {0, 1}n and
an edge from x to y whenever P (y) = x and S(x) = y (S and P stand for “successor
candidate” and “predecessor candidate”). This graph has indegree and outdegree at most
one, and at least one source, namely 0n , so it must have a sink. We seek either a sink,
or a source other than 0n . Thus, PPAD is the class of all total functions whose totality
is proven via the simple combinatorial argument outlined above.
A polynomially computable function f is a polynomial-time reduction from total
search problem P to total search problem Q if, for every input x of P, f (x) is an input
of Q, and furthermore there is another polynomially computable function g such that for
every y ∈ Qf (x) , g(y) ∈ Px . A search problem P in PPAD is called PPAD-complete
if all problems in PPAD reduce to it. Obviously, END OF THE LINE is PPAD-complete;
we now know that 2-NASH is PPAD-complete [6, 4].
In this paper we are interested in succinct games. A succinct game [17] G =
(I, T, U ) is a set of inputs I ∈ P, and two polynomial algorithms T and U . For
each z ∈ I, T (z) returns a type, that is, the number of players r ≤ |z| and an rtuple (t1 , . . . , tr ) where |Sp | = tp . We say that G is of polynomial type if all tp ’s are
bounded by a polynomial in |z|. In this paper we are interested in games of both polynomial (Section 3) and non-polynomial type (Sections 5 and 4). Finally, for any r-tuple
of positive integers s = (s1 , . . . , sr ), where sp ≤ tp , and p ≤ r, U (z, p, s) returns an
integer standing for the utility ups . The game in normal form thus encoded by z ∈ I is
denoted by G(z).
Examples of succinct games (due to space constraints we omit the formal definitions, see [17] for more details) are:
– graphical games [13], where players are nodes on a graph, and the utility of a player
depends only on the strategies of the players in its neighborhood.
– congestion games [7], where strategies are sets of resources, and the utility of a
player is the sum of the delays of the resources in the set it chose, where the delay
is a resource-specific function of the number of players who chose this resource.
– network congestion games, where the strategies of each player are given implicitly
as paths from a source to a sink in a graph; since the number of strategies is potentially exponential, this representation is not of polynomial type; we treat network
congestion games in Section 4.
– multimatrix games where each player plays a different 2-person game with each
other player, and the utilities are added.
– semi-anonymous games (a generalization of symmetric games not considered in
[17]) in which all players have the same set of strategies, and each player has a
utility function that depends solely on the number of other players who choose
each strategy (and not the identities of these players).
– several other classes such as local effect games, scheduling games, hypergraphical
games, network design games, facility location games, etc., as catalogued in [17].
Our main result, shown in the next section, implies that the problem finding a Nash
equilibrium in all of these classes of games can be reduced to 2-player games (equivalently, belongs to the class PPAD).
Lastly, we define a bounded (division-free) straight-line program to be an arithmetic
binary circuit with nodes performing addition, subtraction, or multiplication on their
inputs, or evaluating to pre-set constants, with the additional constraint that the values
of all the nodes remain in [0, 1]. This restriction is not severe, as it can be shown that
an arithmetic circuit of size n with intermediate nodes bounded in absolute value by
2poly(n) can be transformed in polynomial time to fit the above constraint (with the
output scaled down by a factor dependent only on the bound).
3 The Main Result
Given a succinct game, the following problem, called EXPECTED UTILITY, is of interest: Given a mixed strategy profile x1 , . . . , xr , compute the expected utility of player p.
Notice that the result sought is a polynomial in the input variables. It was shown in [17]
that a polynomial-time algorithm for EXPECTED UTILITY (for succinct games of polynomial type) implies a polynomial-time algorithm for computing correlated equilibria
for the succinct game. Here we show a result of a similar flavor.
3.1 Mapping Succinct Games to Graphical Games
Theorem 1. If for a succinct game G of polynomial type there is a bounded divisionfree straight-line program of polynomial length for computing EXPECTED UTILITY,
then G can be mapped in polynomial time to a graphical game G so that there is a
polynomially computable surjective mapping from the set of Nash equilibria of G to the
set of Nash equilibria of G.
Proof. Let G be a succinct game for which there is a bounded straight-line program
for computing EXPECTED UTILITY. In time polynomial in |G|, we will construct a
graphical game G so that the statement of the theorem holds. Suppose that G has r
players, 1, . . . , r, with strategy sets Sp = {1, . . . , tp }, ∀p ≤ r. The players of game
G, which we shall call nodes in the following discussion to distinguish them from the
players of G, will have two strategies each, strategy 0 and strategy 1. We will interpret
the probability with which a node x of G chooses strategy 1 as a real number in [0, 1],
which we will denote, for convenience, by the same symbol x that we use for the node.
Below we describe the nodes of G as well as the role of every node in the construction. We will describe G as a directed network with vertices representing the nodes
(players) of G and directed edges denoting directed flow of information as in [11, 6].
1. For every player p = 1, . . . , r of G and for every pure strategy j ∈ Sp , game G has
a node xpj . Value xpj should be interpreted as the probability with which player p
plays strategy j; in fact, we will establish later that, given a Nash equilibrium of G,
this interpretation yields a Nash equilibrium of G. As we will
in Item 4 below,
Ptsee
p
xpj = 1, ∀p ≤ r.
our construction will ensure that, at any Nash equilibrium, j=1
Therefore, it is legitimate to interpret the set of values {xpj }j as a mixed strategy
for player p in G.
2. For every player p = 1, . . . , r of G and for every pure strategy j ∈ Sp , game G has
p
. The construction of G will ensure that, at a Nash equilibrium,
nodes Ujp and U≤j
p
value Uj equals the utility of player p for playing pure strategy j if every other
player q 6= p plays the mixed strategy specified by the distribution {xqj }j . Also, the
p
construction will ensure that U≤j
= maxj 0 ≤j Ujp0 . Without loss of generality, we
assume that all utilities in G are scaled down to lie in [0, 1].
3. For every node of type Ujp there is a set of nodes in G that simulate the intermediate
variables used by the straight-line program computing the expected utility of player
p for playing pure strategy j when the other players play according to the mixed
strategies specified by {{xqj }j }q6=p . This is possible due to our constraint on the
straight-line program.
4. For every player p of G, there is a set of nodes Ψp defining a component Gp of G
whose
is to guarantee the following at any Nash equilibrium of G:
Ptpurpose
p
p
(a)
j=1 xj = 1
(b) Ujp > Ujp0 =⇒ xpj0 = 0
The structure and the functionality of Gp are described in section 3 of [11], so its
details will be omitted here. Note that the nodes of set Ψp interact only with the
p
p
nodes {Ujp }j , {U≤j
}j and {xpj }j . The nodes of types Ujp and U≤j
are not affected
by the nodes in Ψp and should be interpreted as “input” to Gp , whereas the nodes of
type xpj are only affected by Gp and not by the rest of the game and are the “output”
of Gp . The construction of Gp ensures that they satisfy Properties 4a and 4b.
Having borrowed the construction of the components Gp , p ≤ r, from [11], the only
components of G that remain to be specified are those that compute expected utilities.
With the bound on intermediate variable values, the construction of these components
can be easily done using the games G= , Gζ , G+ , G− , G∗ for assignment, assignment of
a constant ζ, addition, subtraction and multiplication that were defined in [11]. Finally,
p
the components of G that give values to nodes of type U≤j
can be easily constructed
using games Gmax from [11]. It remains to argue that, given a Nash equilibrium of G, we
can find in polynomial time a Nash equilibrium of G and moreover that this mapping is
onto. The first claim follows from the following lemma and the second is easy to verify.
Lemma 1. At a Nash equilibrium of game G, values {{xpj }j }p constitute a Nash equilibrium of game G.
Proof. From the correctness of games Gp , p ≤ r, it follows that, at any Nash equilibPtp p
rium of game G, j=1
xj = 1, ∀p. Moreover, from the correctness of games G= , Gζ ,
G+ , G− , G∗ , it follows that, at any Nash equilibrium of game G, Ujp will be equal to
the utility of player p for playing pure strategy j when every other player q 6= p plays
as specified by the values {xqj }j . From the correctness of Gmax it follows that, at any
p
Nash equilibrium of game G, U≤j
= maxj 0 ≤j Ujp0 , ∀p, j. Finally, from the correctness
of games Gp , p ≤ r, it follows that, at any Nash equilibrium of game G, for every p ≤ r
and for every j, j 0 ∈ Sp , j 6= j 0 : Ujp > Ujp0 =⇒ xpj0 = 0. By combining the above it
follows that {{xpj }j }p constitute a Nash equilibrium of game G. 3.2 Succinct Games in PPAD
We now explore how the mapping described in Theorem 1 can be used in deriving
complexity results for the problem of computing a Nash equilibrium in succinct games.
Theorem 2. If for a succinct game G of polynomial type there is a bounded divisionfree straight-line program of polynomial length for computing EXPECTED UTILITY,
then the problem of computing a Nash equilibrium in the succinct game polynomially
reduces to the problem of computing a Nash equilibrium of a 2-player game.
Proof. We will describe a reduction from the problem of computing a Nash equilibrium in a succinct game to the problem of computing a Nash equilibrium in a graphical
game. This is sufficient since the latter can be reduced to the problem of computing a
Nash equilibrium in a 2-player game [6, 4]. Note that the reduction sought does not follow trivially from Theorem 1; the mapping there makes sure that the exact equilibrium
points of the graphical game can be efficiently mapped to exact equilibrium points of
the succinct game. Here we seek something stronger; we want every approximate Nash
equilibrium of the former to be efficiently mapped to an approximate Nash equilibrium
of the latter. This requirement turns out to be more delicate than the previous one.
Formally, let G be a succinct game for which there is a straight line program for
computing EXPECTED UTILITY and let ε be an accuracy specification. Suppose that G
has r players, 1, . . . , r, with strategy sets Sp = {1, . . . , tp }, ∀p ≤ r. In time polynomial
in |G| + |1/ε|, we will specify a graphical game G and an accuracy ε0 with the property
that, given an ε0 -Nash equilibrium of G, one can recover in polynomial time an ε-Nash
equilibrium of G. In our reduction, the graphical game G will be the same as the one
described in the proof of Theorem 1, while the accuracy specification will be of the
form ε0 = ε/2p(n) , where p(n) is a polynomial in n = |G| that will be be specified
later. Using the same notation for the nodes of game G as we did in Theorem 1, let us
consider if the equivalent of Lemma 1 holds for approximate Nash equilibria.
Observation 1 For any ε0 > 0, there exist ε0 -Nash equilibria of game G in which the
values {{xpj }j }p do not constitute an ε-Nash equilibrium of game G.
Proof. A careful analysis of the mechanics of gadgets Gp , p ≤ r, shows that property (2) which is the defining property of an approximate Nash equilibrium
is not
P
guaranteed to hold. In fact, there are ε0 -equilibria of G in which s∈S−p upjs xs >
P
p
p
0
0
0
s∈S−p uj 0 s xs + ε for some p ≤ r, j and j , and, yet, xj 0 is any value in [0, tp · ε ].
The details are omitted. Moreover, the values {xpj }j do not necessarily constitute a distribution as specified by
the following observation.
Observation
2 For any ε0 > 0, for any p ≤ r, at an ε0 -Nash equilibrium of game G,
P p
j xj is not necessarily equal to 1.
Proof. Again by carefully analyzing the behavior of gadgets Gp , p ≤ r, P
at an ε0 -Nash
equilibrium of game G, it can be shown that there are equilibria in which j xpj can be
any value in 1 ± 2tp ε0 . The details are omitted. Therefore, the extraction of an ε-Nash equilibrium of game G from an ε0 -Nash equilibrium of game G cannot be done by just interpreting the values {xpj } as the probability
distribution of player p. What we show next is that, for the right choice of ε0 , a trim
and renormalize strategy succeeds in deriving an ε-Nash equilibrium of game G from
an ε0 -Nash equilibrium of game G. For any p ≤ r, suppose that {x̂pj }j are the values
than tp ε0 equal to zero (trim)
derived from {xpj }j as follows: make all values
P smaller
p
and renormalize the resulting values so that j x̂j = 1. The argument will rely on the
tightness of the bounds mentioned above, also obtained from the gadgets’ properties:
P
Observation 3 In an ε0 -Nash equilibrium of game G, | j xpj − 1| ≤ 2tp ε0 , and, if
P
P
p
p
p
0
0
s∈S−p ujs xs >
s∈S−p uj 0 s xs + ε , then xj 0 ∈ [0, tp · ε ].
Lemma 2. There exists a polynomial p(n) such that, if ε0 = ε/2p(n) , then, at an ε0 Nash equilibrium of game G, the values {{x̂pj }j }p constitute an ε-Nash equilibrium of
game G.
Proof. We will denote by Ujp (·) the function defined by the straight-line program that
computes the utility of player p for choosing pure strategy j. We need to compare
the values Ujp (x̂) with the values of the nodes Ujp of the graphical game G at an ε0 Nash equilibrium. For convenience, let Ûjp , Ujp (x̂) be the expected utility of player p
for playing pure strategy j when the other players play according to {{x̂qj }j }q6=p . Our
ultimate goal is to show that, at an ε0 -Nash equilibrium of game G, for all p ≤ r, j ≤ tp
Ûjp > Ûjp0 + ε =⇒ x̂pj0 = 0
(3)
Let us take c(n) to be the polynomial bound on 2tp . Using Observation 3, we get that,
for all p, j,
x̂pj (1 − c(n)ε0 ) ≤ xpj ≤ max{c(n)ε0 , x̂pj (1 + c(n)ε0 )}
⇒
x̂pj − c(n)ε0 ≤ xpj ≤ x̂pj + c(n)ε0
(4)
To carry on the analysis, note that, although Ûjp is the output of function Ujp (·) on input
{x̂pj }j,p , Ujp is not the correct output of Ujp (·) on input {xpj }j,p . This is, because, at an ε0 Nash equilibrium of game G, the games that simulate the gates of the arithmetic circuit
introduce an additive error of absolute value up to ε0 per operation. So, to compare
Ujp with Ûjp , we shall compare the “erroneous” evaluation of the arithmetical circuit on
input {xpj }j,p carried inside G against the ideal evaluation of the circuit on input {x̂pj }j,p .
Let us assign a nonnegative “level” to every wire of the arithmetical circuit in the natural
way: the wires to which the input is provided are at level 0 and a wire out of a gate is
at level one plus the maximum level of the gate’s input wires. Since the arithmetical
circuits that compute expected utilities are assumed to be of polynomial length the
maximum level that a wire can be assigned to is q(n), q(·) being some polynomial. The
“erroneous” and the “ideal” evaluations of the circuit on inputs {xpj }j,p and {x̂pj }j,p
respectively satisfy the following property which can be shown by induction:
Lemma 3. Let v, v̂ be the values of a wire at level i of the circuit in the erroneous and
the ideal evaluation respectively. Then
v̂ − g(i)ε0 ≤ v ≤ v̂ + g(i)ε0
where g(i) = 3i · (c(n) + 21 ) − 12 .
By this lemma, the outputs of the two evaluations will satisfy
Ûjp − (2q(n) · (c(n) + 1) − 1)ε0 ≤ Ujp ≤ Ûjp + (2q(n) · (c(n) + 1) − 1)ε0
Thus, setting ε0 =
ε
8c(n)3q(n)
yields |Ujp − Ûjp | ≤ ε/4. After applying the same argument
to Ujp0 and Ûjp0 , we have that Ûjp > Ûjp0 + ε implies Ujp + ε/4 ≥ Ûjp > Ûjp0 + ε ≥
Ujp0 + 3ε/4, and thus Ujp > Ujp0 + ε/2 > Ujp0 + ε0 . Then, from Observation 3, it follows
that xpj0 < tp ε0 and, from the definition of our trimming process, that x̂pj0 = 0. So (3) is
satisfied, therefore making {{x̂pj }j }p an ε-Nash equilibrium. In Section 3.4 we point out that the EXPECTED UTILITY problem in typical succinct
games of polynomial type is very hard. However, in all well known succinct games in
the literature, it turns out that there is a straight-line program of polynomial length that
computes EXPECTED UTILITY:
Corollary 1. The problem of computing a Nash equilibrium in the following families of
succinct games can be polynomially reduced to the same problem for 2-player games:
graphical games, congestion games, multimatrix games, semi-anonymous games, local
effect games, scheduling games, hypergraphical games, network design games, and
facility location games.
Proof. It turns out that, for all these families, there is indeed a straight-line program as
specified in Theorem 2. For graphical games, for example, the program computes explicitly the utility expectation of a player with respect to its neighbors; the other mixed
strategies do not matter. For multimatrix games, the program computes one quadratic
form per constituent game, and adds the expectations (by linearity). For hypergraphical games, the program combines the previous two ideas. For the remaining kinds,
the program combines results of several instances of the following problem (and possibly the two previous ideas, linearity of expectation and explicit expectation calculation):
PnGiven n Bernoulli variables x1 , . . . , xn with Pr[xi = 1] = pi , calculate qj =
Pr[ i=1 xi = j] for j = 0, . . . , n. This can be done by dynamic programming, letting
Pk
k−1
k
qjk = Pr[ i=1 xi = j] (and omitting initializations): qj+1
= (1 − pi )qjk−1 + pi qj−1
,
obviously a polynomial division-free straight-line program. 3.3 An Alternative Proof
We had been looking for some time for an alternative proof of this result, not relying on the machinery of [11]. This proof would start by reducing the Nash equilibrium problem to Brouwer by the reduction of [10]. The Brouwer function in [10] maps
each mixed strategy profile x = (x1 , . . . , xn ) to another (y1 , . . . , yn ), where yi =
arg max (E(x−i ,yi ) [Ui ] − ||yi − xi ||2 ). That is, yi optimizes a trade-off between utility
and distance from xi . It should be possible, by symbolic differentiation of the straightline program, to approximate this optimum and thus the Brouwer function. There are,
though, difficulties in proceeding, because the next step (reduction to Sperner’s Lemma)
seems to require precision incompatible with guarantees obtained this way.
3.4 Intractability
Let us briefly explore the limits of the upper bound in this section.
Proposition 1. There are succinct games of polynomial type for which
UTILITY is #P-hard.
EXPECTED
Proof. Consider the case in which each player has two strategies, true and false,
and the utility of player 1 is 1 if the chosen strategies satisfy a given Boolean formula.
Then the expected utility, when all players play each strategy with probability 21 is the
number of satisfying truth assignments divided by 2n , a #P-hard problem. Thus, the sufficient condition of our Theorem is nontrivial, and there are games of
polynomial type that do not satisfy it. Are there games of polynomial type for which
computing Nash equilibria is intractable beyond PPAD? This is an important open
question. Naturally, computing a Nash equilibrium of a general succinct game is EXPhard (recall that it is so even for 2-person zero-sum games [8, 9], and the nonzero version can be easily seen to be complete for the exponential counterpart of PPAD).
Finally, it is interesting to ask whether our sufficient condition (polynomial computability of EXPECTED UTILITY by a bounded division-free straight-line program) is
strictly weaker than the condition in [17] for correlated equilibria (polynomial computability of EXPECTED UTILITY by Turing machines). It turns out1 that it is, unless
⊕P is in nonuniform polynomial time [2]. Determining the precise complexity nature
of this condition is another interesting open problem.
1
Many thanks to Peter Bürgisser for pointing this out to us.
4 Network Congestion Games
A network congestion game [7] is specified by a network with delay functions, that is,
a directed graph (V, E) with a pair of nodes (ap , bp ) for each player p, and also, for
each edge e ∈ E, a delay function de mapping [n] to the positive integers; for each
possible number of players “using” edge e, de assigns a delay. The set of strategies for
player p is the set of all paths from ap to bp . Finally, the payoffs are determined as
follows: If s = (s1 , . . . , sn ) is a pure strategy profile, define ce (s) = |{p : e ∈ sp }|
(here
P we consider paths as sets of edges); then the utility of player p under s is simply
− e∈sp de (ce (s)), the negation of the total delay on the edges in p’s strategy. It was
shown in [7] that a pure Nash equilibrium of a network congestion game (known to
always exist) can be found in polynomial time when the game is symmetric (ap = a1
and bp = b1 for all p), and PLS-complete in the general case. There is no known
polynomial-time algorithm for finding Nash equilibria (or any kind of equilibria, such
as correlated [17]) in general network congestion games. We prove:
Theorem 3. The problem of computing a Nash equilibrium of a network congestion
game polynomially reduces to the problem of computing a Nash equilibrium of a 2player game.
Proof. (Sketch.) We will map a network congestion game to a graphical game G. To
finish the proof one needs to use techniques parallel to Section 3.2. To simulate network
congestion games by graphical games we use a nonstandard representation of mixed
strategy: We consider a mixed strategy for player p to be a unit flow from ap to bp , that
is, an assignment of nonnegative values fp (e) to the edges of the network such that all
nodes are balanced except for ap who has a deficit of 1 and bp who has a gain of 1.
Intuitively, fp (e) corresponds to the sum of the probabilities of all paths that use e.
It turns out that such flow can be set up in the simulating graphical game by a gadget
similar to the one that sets up the mixed strategy of each player. In particular, for every
player p and for every edge e of the network there will be a player in the graphical game
whose value will represent fp (e). Moreover, for every node v 6= ap , bp of the network,
there will be a player Svp in the graphical game whose value will be equal to the sum of
the flows of player p on the edges entering node v; there will also be a gadget Gvp similar
to the one used in proof of Theorem 1, whose purpose will be to distribute the flow of
player p entering v, i.e. value Svp , to the edges leaving node v, therefore guaranteeing
that Kirchhoff’s first law holds. The distribution of the value Svp on the edges leaving
v will be determined by finding the net delays between their endpoints and node bp as
specified by the next paragraphs. Finally, note that the gadgets for nodes ap and bp are
similar but will inject a gain of 1 at ap and a deficit of 1 at bp . Some scaling will be
needed to make sure that all computed values are in [0, 1].
The rest of the construction is based on the following Lemma, whose simple proof
we omit. Fix a player p and a set of unit flows fq for the other players. These induce an
expected delay on each edge e, E[de (ce (k))] where k is 1 (for player p) plus the sum of
n − 1 variables that are 1 with probability fq (e) and else 0. Call this quantity Dp (e).
Lemma 4. A set of unit flows fp (e), p = 1, . . . , n is an ε-Nash equilibrium if and only
if fp (e) > 0 implies that e lies on a path whose length (defined as net delay under Dp ),
is at most ε above the length of the shortest path from ap to bp .
We shall show that these conditions can be calculated by a straight-line program in
polynomial time; this implies the Theorem. This is done as follows: First we compute
the distances Dp (e) for all edges and players by dynamic programming, as in the proof
of Corollary 1. Then, for each player p and edge (u, v) we calculate the shortest path
distances, under Dp , (a) from ap to bp ; (b) from ap to u and (c) from v to bp . This is done
by the Bellman-Ford algorithm, which is a straight line program with the additional use
of the min operator (see [11] for gadget). The condition then requires that the sum of
the latter two and Dp (u, v) be at most the former plus ε. This completes the proof. 5 Extensive Form Games
An r-player extensive form game (see, e.g., [16]) is represented by a game tree with
each non-leaf vertex v assigned to a player p(v), who “plays” by choosing one of the
outgoing labeled edges, and with a vector of payoffs upx at each leaf x (let X be the
set of leaves). All edges have labels, with the constraint that l(v, v 0 ) 6= l(v, v 00 ). The
vertex set is partitioned into information sets I ∈ I, with all v ∈ I owned by the same
player p(I), and having identical sets of outgoing edge labels LI . We also define Ip =
{I ∈ I|p(I) = p}. Information sets represent a player’s knowledge of the game state.
A behavioral strategy σ p for player p is an assignment of distributions {σjp,I }j∈LI over
the outgoing edge labels of each I ∈ Ip . A behavioral strategy profile σ = (σ 1 , . . . , σ r )
induces a distribution over the leaves of the game tree, and hence expected utilities. A
behavioral Nash equilibrium is the natural equivalent of the normal form’s mixed Nash
equilibrium: a σ such that no player p can change σ p and increase his expected payoff.
Theorem 4. The problem of computing a behavioral Nash equilibrium (and, in fact,
a subgame perfect equilibrium [16]) in an extensive form game Γ is polynomially reducible to computing a mixed Nash equilibrium in a 2-player normal form game.
Proof. (Sketch.) As in Section 4, we will map an extensive form congestion game to a
graphical game, and omit the rest of the argument, which is also akin to Section 3.2. The
graphical game construction is similar to that in Section 3.1. Using nodes with strategy
sets {0, 1},
1. For every information set I with p(I) = p and an outgoing edge label j ∈ LI ,
make a node σjp,I , to represent the probability of picking j.
2. For every information set I and every j ∈ LI make a node UjI ; the value of UjI will
represent the utility of player p(I) resulting from the optimal choice of distributions
player p(I) can make in the part of the tree below information set I given that the
player arrived at information set I and chose j and assuming that the other players
0
play as prescribed by the values {σjq,I }q=p(I 0 )6=p ; the weighting of the vertices of
I when computing UjI is defined by the probabilities of the other players on the
edges that connect I to the closest information set of p(I) above I. Let U I be the
0
maximum over UjI . Assuming the values U I for the information sets I 0 below I
are computed, value UjI can be found by arithmetic operations.
3. Finally, for every information set, take a gadget GI similar to Gp above that guaranP
p(I),I
p(I),I
tees that (i) j∈LI σj
= 0.
= 1, and (ii) UjI > UjI0 =⇒ σj 0
Further details are omitted. The construction works by arguments parallel to the proof
of Theorem 1. References
1. R. J. Aumann. “Subjectivity and Correlation in Randomized Strategies,” Journal of Mathematical Economics, 1 , pp. 67-95, 1974.
2. P. Bürgisser. “On the structure of Valiant’s complexity classes,” Discr. Math. Theoret. Comp.
Sci., 3, pp. 73–94, 1999.
3. X. Chen and X. Deng. “3-NASH is PPAD-Complete,” ECCC, TR05-134, 2005.
4. X. Chen and X. Deng. “Settling the Complexity of 2-Player Nash-Equilibrium,” ECCC,
TR05-140, 2005.
5. C. Daskalakis and C. H. Papadimitriou. “Three-Player Games Are Hard,” ECCC, TR05-139,
2005.
6. C. Daskalakis, P. W. Goldberg and C. H. Papadimitriou. “The Complexity of Computing a
Nash Equilibrium,” Proceedings of 38th STOC, 2006.
7. A. Fabrikant, C.H. Papadimitriou and K. Talwar. “The Complexity of Pure Nash Equilibria,”
Proceedings of 36th STOC, 2004.
8. J. Feigenbaum, D. Koller and P. Shor. “A game-theoretic classification of interactive complexity classes,” IEEE Conference on Structure in Complexity Theory, 1995.
9. L. Fortnow, R. Impagliazzo, V. Kabanets and C. Umans. “On the Complexity of Succinct
Zero-Sum Games,” IEEE Conference on Computational Complexity, 2005.
10. J. Geanakoplos. “Nash and Walras Equilibrium via Brouwer,” Economic Theory, 21, 2003.
11. P. W. Goldberg and C. H. Papadimitriou. “Reducibility Among Equilibrium Problems,” Proceedings of 38th STOC, 2006.
12. D. S. Johnson, C. H. Papadimitriou and M. Yannakakis, “How Easy is Local Search?,” J.
Comput. Syst. Sci. 37, 1, pp. 79–100,1988.
13. M. Kearns, M. Littman and S. Singh. “Graphical Models for Game Theory,” In UAI, 2001.
14. C. E. Lemke and J. T. Howson Jr.. “Equilibrium points of bimatrix games,” Journal of the
Society for Industrial and Applied Mathematics, 12, 413-423, 1964.
15. J. Nash. “Noncooperative Games,” Annals of Mathematics, 54, 289-295, 1951.
16. M.J. Osborne and A. Rubinstein. A Course in Game Theory, MIT Press, 1994.
17. C. H. Papadimitriou. “Computing Correlated Equilibria in Multiplayer Games,” STOC, 2005.
18. C. H. Papadimitriou. “On the Complexity of the Parity Argument and Other Inefficient Proofs
of Existence,” J. Comput. Syst. Sci. 48, 3, pp. 498–532, 1994.
19. R. Savani and B. von Stengel. “Exponentially many steps for finding a Nash equilibrium in
a Bimatrix Game”. Proceedings of 45th FOCS, 2004.
20. B. von Stengel, Computing equilibria for two-person games. In R.J. Aumann and S. Hart, editors, Handbook of Game Theory, Vol. 3, pp. 1723–1759. North-Holland, Amsterdam, 2002.